Caideyipi commented on code in PR #18663:
URL: https://github.com/apache/iotdb/pull/18663#discussion_r4033082977
##########
iotdb-core/confignode/src/main/java/org/apache/iotdb/confignode/manager/load/balancer/RegionBalancer.java:
##########
@@ -118,6 +130,14 @@ public CreateRegionGroupsPlan
genRegionGroupsAllocationPlan(
final int allotment = entry.getValue();
final int replicationFactor =
getClusterSchemaManager().getReplicationFactor(database,
consensusGroupType);
+ final Integer requestPreferredDataNode =
+ preferredDataNodeMap == null ? null :
preferredDataNodeMap.get(database);
+ final int preferredDataNodeId =
+ TConsensusGroupType.DataRegion.equals(consensusGroupType)
+ ? (requestPreferredDataNode != null
+ ? requestPreferredDataNode
+ : getPreferredDataNodeId(database))
Review Comment:
[P1] Keep the DataNode preference for a single LOAD out of the persistent
database schema
When a request has no hint, this code falls back to the
`preferredDataNodeId` persisted in the database schema. There is no mechanism
to clear this field, so after a database is created by LOAD, later DataRegion
expansion triggered by ordinary INSERTs, repair tasks, etc. will continue to
favor the original DataNode. Even if `load_tsfile_prefer_local_node=false` is
applied at runtime, `PartitionManager` still passes an empty map and this
fallback remains active. With RF=1, subsequent Regions may even keep
accumulating on the same node. Please keep the hint scoped to the request, or
clear it after its first use.
##########
iotdb-core/datanode/src/main/java/org/apache/iotdb/db/queryengine/plan/analyze/ClusterPartitionFetcher.java:
##########
@@ -319,6 +319,14 @@ public DataPartition getOrCreateDataPartition(
@Override
public DataPartition getOrCreateDataPartition(
final List<DataPartitionQueryParam> dataPartitionQueryParams, final
String userName) {
+ return getOrCreateDataPartition(dataPartitionQueryParams, userName, -1);
Review Comment:
[P2] Tablet conversion paths lose the preferred DataNode hint
The existing two-argument overload is forwarded here with
`preferredDataNodeId=-1`. However, small TsFiles are converted directly to
`InsertTablet` before normal schema analysis, and conversion for type
mismatches also uses the regular INSERT analysis API. These paths ultimately
call that two-argument overload, so the local node will not be preferred even
when LOAD creates a database or DataRegion. Please propagate the preferred ID
into the partition requests for the converted tree-model and table-model
inserts.
##########
iotdb-protocol/thrift-confignode/src/main/thrift/confignode.thrift:
##########
@@ -241,6 +241,7 @@ struct TDatabaseSchema {
10: optional i64 timePartitionOrigin
11: optional bool isTableModel
12: optional bool needLastCache
+ 13: optional i32 preferredDataNodeId
Review Comment:
[P2] Do not propagate a cluster-local DataNodeId through Pipe in the
database schema
`TDatabaseSchema` is sent downstream through ConfigRegion Pipe as part of a
`CreateDatabase` plan. The receiver replaces sender-local settings such as the
replication factor and time partition configuration, but does not clear this
field. DataNodeId is a cluster-local identifier. If a node with the same ID
exists in the target cluster, subsequent DataRegions will incorrectly favor
that node. Disabling `load_tsfile_prefer_local_node` on the target cluster does
not prevent the schema fallback either. If this field is retained, the Pipe
receiver should at least call `unsetPreferredDataNodeId()`.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]